Hi,
First of all, thank you for your excellent work. I successfully reproduced the results on MOSEI and SIMS, but I encountered some problems when trying to reproduce the MOSI results with SmolLM2-1.7B
During training, I met several issues, for example the mismatched tensor shapes between input_ids and attention_mask inside msaLM (particularly between the second and third forward printouts).
I managed to fix some of them, but I still cannot reproduce the performance reported in the paper — my current results are worse.
Could you please share the complete configuration files for MOSI with SmolLM2-1.7B, including both the DeepMLF config, and the BiEnc config (if applicable)?
Thank you very much for your time and for releasing such an interesting framework!
Hi,
First of all, thank you for your excellent work. I successfully reproduced the results on MOSEI and SIMS, but I encountered some problems when trying to reproduce the MOSI results with SmolLM2-1.7B
During training, I met several issues, for example the mismatched tensor shapes between input_ids and attention_mask inside msaLM (particularly between the second and third forward printouts).
I managed to fix some of them, but I still cannot reproduce the performance reported in the paper — my current results are worse.
Could you please share the complete configuration files for MOSI with SmolLM2-1.7B, including both the DeepMLF config, and the BiEnc config (if applicable)?
Thank you very much for your time and for releasing such an interesting framework!